Research
Research
Embedded Files
Human motion understanding
Human motion understanding
The capability of understanding biological motion - meaning motion produced by a living being - is a trigger for developing high-level perceptual tasks and the ability to socially interact. In our research we are interested in the study and development of computational models for human motion representation and recognition inspired by models of motion perception in humans and how it develops in childhood.
Specific tasks of interest are the following:
- Designing spatio-temporal multi-scale motion descriptors
- Identifying motion primitives
- Detecting action hierarchies and categories
- Exploring multi-view motion representations
Refs.
A. Vignolo, N. Noceti, F. Rea, A. Sciutti, F. Odone, G. Sandini. Detecting biological motion for human-robot interaction: a link between perception and action. Frontiers in Robotics and AI. 2017
N.Noceti, A.Sciutti, F.Odone, G.Sandini. Exploring biological motion regularities of human actions: a new perspective on video analysis. ACM Transactions on Applied Perception. 2017
A. Vignolo, N.Noceti, A.Sciutti, F. Rea, F. Odone, G.Sandini. The Complexity of Biological Motion: A Temporal Multi- Resolution Motion Descriptor for Human Detection in Videos. IEEE International Conference on Developmental Learning and Epigenetic Robotics. 2016
High-Performance Computing for Computer Vision
High-Performance Computing for Computer Vision
The earliest task of motion analysis is always the ability to identify the moving regions in image streams, often referred to as motion-based image (and video) segmentation. In these fields, most of the work in the recent decades has a focus on accuracy of results; only a marginal share of the literature deals with performance intended as execution time of the algorithms. This latter aspect, however, plays a critical role in real-life visual applications.
We are interested in devising methods leveraging modern highly-parallel computer architectures (e.g. many-core processors) and throughput-oriented devices (e.g. GPUs) using state-of-the-art numerical methods for background subtraction and optical flow estimation, specifically intended to couple effectiveness and computational performance.
Visual computing for Ambient Assisted Living
Visual computing for Ambient Assisted Living
Recent years have witnessed an increasing interest in Ambient Assisted Living (AAL) technologies specifically targeted to elderly, a share of world population in continuous growing. In our research, we propose a paradigm shift in assistance, in which intelligent environments act as personalized, social-aware and evolving cognitive prostheses to assisted people.
We are interested in particular in the development and validation of visual computing methods to evaluate the quantity and quality of people activities, representing the basis for evaluations that can complement the common clinical protocols adopted by geriatricians. Specific tasks of interest are
- Using RGB-D data streams to evaluate the motility of patients
- Video-based fine-grained scene segmentation
- Pose classification from skeleton representations
- Recognition of Activities of Daily Living (ADL)
Refs.
Chiara Martini, Nicoletta Noceti, Manuela Chessa, Annalisa Barla, Francesca Odone, Alessandro Verri, Alberto Pilotto, Alberto Cella and Gian Andrea Rollandi. An integrated approach for estimating the frailty index in the elderly. VISAPP 2018
Manuela Chessa, Nicoletta Noceti, Chiara Martini, Fabio Solari and Francesca Odone. Design of assistive tools for the market. Assistive Computer Vision (Elsevier Book)
Machine Learning for time series analysis
Machine Learning for time series analysis
In the age of digital data, a large number of real world problems can be casted into the general framework of time series analysis. In industry, for instance, they may describe the life cycle of a sensorized component; in financial contexts, they may be connected with the evolution of a certain economic phenomenon.
A common trait of all these scenarios is the necessity of understanding the presence of prototypical temporal behaviors of the data and the consequent ability to identify possibly anomalous configurations.
In these contexts, we are interested in Machine Learning methodologies specifically tailored for streams of temporal data, with particular emphasis on
- Developing supervised and unsupervised methodologies for learning patterns from time data
- Designing strategies to build incremental models over time
The outcomes of our research are applied to different scenarios, ranging from surveillance and monitoring (to learn patterns of activities), to industrial applications (for maintenance and early anomaly detection), to financial services (for bank services, or in combination with Block-Chain technologies).
Refs.
N.Noceti, F.Odone. A prototype application for long-time behaviour modelling and abnormal events detection. VISAPP 2016
N.Noceti, F. Odone. Learning common behaviors from large sets of unlabeled temporal series. Image and Vision Computing, 30(11), 875-895. 2012
Google Sites
Report abuse